Method and system for nonlinear compensation of ultra-high order qam signal based on time feature memory neural network

Abstract

The present application relates to the technical field of ultra-high order QAM signal nonlinear compensation, in particular to an ultra-high order QAM signal nonlinear compensation method and system based on a time feature memory neural network, the method comprising: S1: under different optical signal-to-noise ratios, a plurality of quadrature amplitude modulation signals are obtained, the quadrature amplitude modulation signals are subjected to zero-mean normalization processing, and preprocessed quadrature amplitude modulation signals are obtained; S2: the preprocessed quadrature amplitude modulation signals are divided into a training set and a test set, a signal nonlinear compensation model is constructed, the training set is used to train the signal nonlinear compensation model, and a trained signal nonlinear compensation model is obtained; and S3: based on the trained signal nonlinear compensation model, nonlinear compensation is performed on the quadrature amplitude modulation signals to be compensated. The present application can extract the time sequence features of the signals and capture the bidirectional time dependence, thereby achieving the purpose of nonlinear distortion compensation.

Description

Technical Field

[0001] The present invention relates to the technical field of ultra-high-order QAM signal nonlinear compensation, and in particular to an ultra-high-order QAM signal nonlinear compensation method and system based on a time characteristic memory neural network. Background Art

[0002] According to data from Cisco's authoritative "2020 Internet Report," the compound annual growth rate (CAGR) for internet users will remain stable at approximately 6% between 2018 and 2023, indicating a continued surge in network traffic and a more pressing demand for data transmission efficiency. Against this backdrop, coherent optical communication technology, with its unique advantages, is emerging as a key enabler of addressing this traffic growth challenge.

[0003] To effectively improve spectral efficiency (SE), one common strategy adopted by the industry is to increase the order of signal modulation. This approach has been fully validated in experiments involving probability shaping (PS) using 1024-QAM and even 4096-QAM. However, the transmission of ultra-high-order signals often involves complex hardware deployment and high costs, which undoubtedly poses a significant obstacle to its widespread application. Therefore, advanced digital signal processing (DSP) algorithms are particularly important in compensating for signal impairments and optimizing transmission performance.

[0004] Achieving effective equalization and noise suppression for ultra-high-order Quadrature Amplitude Modulation (QAM) signals, particularly nonlinear compensation (NLC), has become a critical challenge that needs to be overcome. The nonlinear characteristics of optical fiber, caused by the Kerr effect, impose an inherent nonlinear Shannon limit on the information transmission capacity of optical fiber communication systems. This problem becomes even more pronounced when it interacts with the spontaneous emission noise generated by optical amplifiers. Furthermore, as the order of QAM signals increases, the points on the constellation become increasingly dense, significantly reducing the Euclidean distance between points. This significantly reduces the resistance of high-order signals to noise and linear / nonlinear impairments, requiring a higher optical signal-to-noise ratio (OSNR) to ensure transmission quality.

[0005] Currently, mainstream nonlinear compensation technologies include digital backpropagation (DBP), Volterra series-based nonlinear equalizer (VNLE), and phase conjugate double wave (PCTW). Although DBP can address impairments by solving the backpropagation nonlinear Schrödinger equation (NLSE), its high computational complexity limits its commercial feasibility. VNLE utilizes Volterra series transfer functions for modeling, but also faces the challenges of massive computational resource consumption and difficulty in handling high-order nonlinear effects. Although PCTW technology can effectively mitigate first-order nonlinear effects through digital coherent superposition, it unfortunately sacrifices half of the spectral efficiency.

[0006] Against this backdrop, neural networks, with their exceptional fitting and analysis capabilities, have demonstrated tremendous potential in the field of nonlinear equalization in optical fiber communication systems. In recent years, machine learning (ML) technology has garnered significant attention for addressing nonlinear impairments in coherent optical communications. From the early artificial neural networks (ANNs) and convolutional neural networks (CNNs) used for 16-QAM to the recurrent neural networks (RNNs) used for 64-QAM, significant progress has been made in practical applications. Of particular note is the successful application of long short-term memory (LSTM) networks in mitigating impairments in 16-QAM coherent optical transmission systems, further demonstrating the promising potential of LSTM-based post-processing techniques in addressing such complex challenges.

[0007] However, it's worth noting that previous research has largely focused on low-order modulated signals, often requiring additional linear equalizers for preprocessing before the neural network architecture. For ultra-high-order signals, the drastic reduction in Euclidean distance between constellation points leads to a sharp increase in the signal's sensitivity to linear and nonlinear impairments, undoubtedly posing a more demanding challenge to the application of neural networks. Even minor distortions can cause rapid signal attenuation, and research on the application of neural networks to such high-order signals is still in its infancy, with ample room for exploration. Summary of the Invention

[0008] To solve the above technical problems, the present invention provides a nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network, comprising the following steps:

[0009] S1: Under different optical signal-to-noise ratios, obtaining multiple groups of quadrature amplitude modulation signals, performing zero-mean normalization processing on the quadrature amplitude modulation signals, and obtaining pre-processed quadrature amplitude modulation signals;

[0010] S2: Dividing the preprocessed quadrature amplitude modulation signal into a training set and a test set, constructing a signal nonlinear compensation model, and training the signal nonlinear compensation model using the training set to obtain a trained signal nonlinear compensation model;

[0011] S3: Based on the trained signal nonlinear compensation model, perform nonlinear compensation on the quadrature amplitude modulation signal to be compensated.

[0012] In one embodiment of the present invention, the signal nonlinear compensation model includes an input layer, a temporal convolutional network layer, a bidirectional long short-term memory neural network layer, and an output layer;

[0013] The input layer processes the training set to obtain a modulated signal in the form of an embedded vector;

[0014] The time sequence convolution network layer extracts features of the modulated signal in the form of the embedding vector to obtain feature sequence data;

[0015] The bidirectional long short-term memory neural network layer processes the feature sequence data in a forward direction and a reverse direction respectively to capture bidirectional time dependence relationship therein and obtain compensated feature sequence data.

[0016] The output layer outputs the compensated feature sequence data.

[0017] In an embodiment of the present application, the time sequence convolution network layer is configured to include a plurality of residual units connected in sequence, each residual unit including a first branch and a second branch, the first branch and the second branch extracting features of the modulated signal in the form of the embedding vector, and performing a concatenation operation on output features of the two branches to obtain the feature sequence data.

[0018] In an embodiment of the present application, the first branch includes dilated convolution layers, layer batch normalization processing layers, activation layers, Dropout layers, dilated convolution layers, layer batch normalization processing layers, activation layers and Dropout layers connected in sequence.

[0019] In an embodiment of the present application, an expression of a processing process of the dilated convolution layer is as follows:

[0020]

[0021] wherein s represents an element of an input sequence x, d is a dilated factor, k is a size of a filter kernel, s-d·i represents a past direction, and f is a filter set.

[0022] In an embodiment of the present application, the activation layer uses a Tanh function.

[0023] In an embodiment of the present application, the second branch is a residual skip branch including a 1×1 point-wise convolution layer.

[0024] In an embodiment of the present application, the bidirectional long short-term memory neural network layer includes a forward long short-term memory network unit, a reverse long short-term memory network unit and a concatenation operation layer.

[0025] The forward long short-term memory network unit and the reverse long short-term memory network unit process the feature sequence data in a forward direction and a reverse direction respectively to obtain a forward feature sequence and a reverse feature sequence, and the concatenation operation layer performs a concatenation operation on the forward feature sequence and the reverse feature sequence to obtain the compensated feature sequence data.

[0026] Based on the same inventive concept, the present invention also provides a nonlinear compensation system for ultra-high-order QAM signals based on a time-feature memory neural network, which includes the following modules:

[0027] A data acquisition and preprocessing module is used to acquire multiple groups of orthogonal amplitude modulation signals under different optical signal-to-noise ratios, perform zero-mean normalization processing on the orthogonal amplitude modulation signals, and obtain preprocessed orthogonal amplitude modulation signals;

[0028] A model construction and training module is used to divide the preprocessed orthogonal amplitude modulation signal into a training set and a test set, construct a signal nonlinear compensation model, and train the signal nonlinear compensation model using the training set to obtain a trained signal nonlinear compensation model;

[0029] The signal nonlinear compensation module is used to perform nonlinear compensation on the quadrature amplitude modulation signal to be compensated based on the trained signal nonlinear compensation model.

[0030] The present invention also provides a computer storage medium, which stores a computer software product. The computer software product includes several instructions for enabling a computer device to execute the instructions of the ultra-high-order QAM signal nonlinear compensation method based on time characteristic memory neural network.

[0031] The above technical solution of the present invention has the following advantages over the prior art:

[0032] 1. By constructing a time feature memory neural network model, the present invention can efficiently capture and compensate for the nonlinear distortion in ultra-high-order QAM signals. This is due to the combined effect of the timing convolutional network layer and the bidirectional long short-term memory neural network layer in the model. They can respectively extract the timing features of the signal and capture the bidirectional time dependency, thereby improving the accuracy and efficiency of compensation.

[0033] 2. The temporal convolutional network layer contains multiple residual units. Each unit, through the combination of the first and second branches, can extract signal features from multiple levels and angles. In particular, the combination of the dilated convolution layer and the 1×1 point-by-point convolution layer in the first branch increases the network's receptive field while maintaining feature diversity, helping to improve the model's generalization capabilities.

[0034] 3. The introduction of a bidirectional long short-term memory neural network layer enables the model to process both forward and reverse feature sequence data simultaneously, thereby more comprehensively capturing the temporal dependencies in the signal. This bidirectional processing mechanism helps improve the model's understanding and ability to compensate for the dynamic characteristics of the signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0036] Figure 1 This is a flow chart of a nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network provided in the first embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the signal nonlinear compensation model structure based on the time feature memory neural network (TFM NN);

[0038] Figure 3 It is a schematic diagram of the temporal convolutional network layer structure;

[0039] Figure 4 This is a diagram of the experimental setup for transmitting PS-PDM1024 / 4096-QAM signals over 80 km of standard single-mode fiber (SSMF);

[0040] Figure 5 (a) and (c) show the NGMI performance measured with / without the TFMNN equalizer in a back-to-back (BTB) transmission experiment, respectively; (b) and (d) show the NGMI performance measured with the traditional VE and the proposed TFMNN in an 80 km standard single-mode fiber (SSMF) transmission experiment, respectively;

[0041] Figure 6 Schematic diagram of the structure of the ultra-high-order QAM signal nonlinear compensation system based on the time characteristic memory neural network provided in the second embodiment of the present invention;

[0042] Description of the accompanying figures in the specification: 100, data acquisition and preprocessing module; 200, model construction and training module; 300, signal nonlinear compensation module. DETAILED DESCRIPTION

[0043] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0044] Example 1

[0045] Reference Figure 1 As shown, the present invention provides a nonlinear compensation method for ultra-high-order QAM signals based on a time feature memory neural network, the method comprising the following steps:

[0046] S1: Obtain a plurality of sets of quadrature amplitude modulation (QAM) signals under different optical signal-to-noise ratios (OSNRs), each set of quadrature amplitude modulation signals including real and imaginary components of two polarization states of 10,000 symbols, and recording information of the current symbol and the surrounding environment;

[0047] Perform zero-mean normalization processing on the quadrature amplitude modulation signals to obtain preprocessed quadrature amplitude modulation signals, the preprocessed quadrature amplitude modulation signals being pseudo-random binary sequences (PRBSs);

[0048] S2: Divide the preprocessed quadrature amplitude modulation signals into a training set and a test set, construct a time feature memory (TFM) neural network including an input layer, a time convolution network (TCN) layer, a bidirectional long short-term memory (Bi-LSTM) neural network layer, and an output layer as a signal nonlinear compensation model, train the signal nonlinear compensation model using the training set, monitor the learning progress of the model by calculating the mean square error (MSE) once every 25 epochs, and set an early stopping mechanism, that is, if the MSE does not significantly improve after 250 consecutive epochs, the training is terminated in advance to prevent overfitting, and obtain a trained signal nonlinear compensation model;

[0049] S3: Perform nonlinear compensation on the quadrature amplitude modulation signals to be compensated based on the trained signal nonlinear compensation model.

[0050] As can be seen from the above technical solutions, in step S1, the present application considers quadrature amplitude modulation (QAM) signals under different OSNR conditions, ensuring that the model can process signals in various noise environments. Each set of QAM signals contains real and imaginary components of two polarization states, as well as information of the current symbol and its surrounding environment, which helps the model more accurately capture the dynamic characteristics and time dependence of the signal.

[0051] In step S2, the time feature memory neural network uses the combined action of the time convolution network layer and the bidirectional long short-term memory neural network layer to extract the time sequence features of the signal and capture the bidirectional time dependence, respectively, which plays a crucial role in improving the accuracy and efficiency of signal impairment compensation.

[0052] In step S3, the trained model is used to perform nonlinear compensation on the QAM signals to be compensated, which can significantly reduce the nonlinear distortion in the signal and improve the transmission quality and spectral efficiency of the signal.

[0053] As shown in Figure 2 the signal nonlinear compensation model is a time feature memory neural network, specifically including an input layer, a time convolution network (TCN) layer, a bidirectional long short-term memory (Bi-LSTM) neural network layer, and an output layer.

[0054] The input layer processes the training set to obtain a modulated signal in the form of an embedded vector;

[0055] The temporal convolutional network layer extracts features from the modulation signal in the embedded vector form to obtain feature sequence data;

[0056] The bidirectional long short-term memory neural network layer processes the feature sequence data in a forward and reverse order respectively, captures the bidirectional time dependency therein, and obtains compensated feature sequence data;

[0057] The output layer outputs the compensated feature sequence data.

[0058] Furthermore, the temporal convolutional network layer is configured to include: a plurality of residual units connected in sequence, each residual unit including a first branch and a second branch, the first branch and the second branch respectively extract features of the modulated signal in the form of the embedded vector, and perform a splicing operation on the output features of the two branches to obtain the feature sequence data.

[0059] Specifically, the first branch includes a dilated convolutional layer, a batch normalization layer, an activation layer, a dropout layer, a dilated convolutional layer, a batch normalization layer, an activation layer, and a dropout layer connected in sequence. Preferably, the activation layer uses a Tanh function. The second branch is a residual skip branch including a 1×1 point-by-point convolutional layer.

[0060] The expression of the data processing process of the dilated convolution layer is as follows:

[0061]

[0062] Where s represents the element of the input sequence x, d is the dilation factor, k is the size of the filter kernel, sd·i represents the past direction, and f is the filter set. When d=1, it simplifies to a regular convolution. Larger dilation can expand the receptive field without requiring a lot of computational cost. By increasing the filter kernel size k or the dilation factor d, the receptive field of TCN will be expanded. For example, when the convolution kernel is 3 and the dilation factor d=o(2 i ) grows exponentially, TCNs can cover all input values. Each layer involves a residual block to the next, enhancing the modeling of long-range dependencies. Therefore, TCN's dilated convolutions and residual blocks can effectively handle long-range dependencies, making it ideal for tasks such as digital signal processing in coherent optical transmission.

[0063] Figure 3A dilated convolution with a kernel size of 3 and a dilation factor of d=n is demonstrated. The receptive field of the TCN architecture is determined by the network depth n, the filter kernel size k, and the dilation factor d. To ensure the stability of deep TCNs and optimize feature extraction, residual layers are employed, consisting of two causal convolutional layers and a nonlinear activation layer. Batch normalization is applied to the post-convolutional layers to maintain stability, while dropout layers help mitigate overfitting. To match the input and output dimensions, 1×1 convolutions are used to properly align the tensors before element-wise addition. Weight normalization is applied to each convolutional layer in the residual block to prevent exploding gradients. To enhance generalization and reduce overfitting, L2 regularization is incorporated into the TCN architecture to constrain the L2 norm of the parameters. Each residual block consists of two convolutional layers with a stride of 1. The residual network architecture includes shortcut connections that bypass certain layers and pass raw data directly to the next layer. These additional connections do not increase the complexity of the model.

[0064] Figure 3 This paper visually demonstrates a dilated convolution operation with a specific configuration, where the kernel size is 3 and the dilation factor d is set equal to the network depth n. In the TCN (Temporal Convolutional Network) architecture, the size of the receptive field is determined by several key factors, including the network depth n, the kernel size k, and the dilation factor d. To enhance the stability of deep TCNs and optimize their feature extraction capabilities, residual layers are used. These residual layers consist of two causal convolutional layers with a nonlinear activation layer between them.

[0065] To maintain model stability, we apply batch normalization after the convolution operation. Furthermore, we introduce a dropout layer to mitigate the risk of overfitting. To ensure that the input and output dimensions match, we strategically use 1×1 convolutions to achieve the necessary tensor alignment before element-wise addition.

[0066] Within the residual block, each convolutional layer is additionally equipped with a weight normalization mechanism, a design that aims to prevent the gradient explosion problem. To further improve the generalization ability of the model and reduce the risk of overfitting, an L2 regularization strategy is integrated into the TCN structure to achieve this goal by constraining the L2 norm of the model parameters. It is worth noting that each residual block is composed of two convolutional layers with a span of 1. In addition, the residual network architecture also cleverly incorporates shortcut connections (SkipConnections), which allow the original data to bypass certain layers and pass directly to the next layer. This design not only enhances the flow of information, but also does not increase the computational complexity of the model.

[0067] In summary, through this series of carefully designed components and strategies, the TCN architecture maintains high efficiency while significantly improving its stability and generalization capabilities.

[0068] By utilizing the TCN layer for feature extraction, the bidirectional long short-term memory (Bi-LSTM) neural network layer can more effectively process complex ultra-high-order QAM signals with a simple structure. The Bi-LSTM network retains the advantages of LSTM, effectively capturing temporal correlations, and integrates a gating mechanism to manage information flow, thereby solving the vanishing gradient problem. Its bidirectionality enhances the integration of information from past and future symbols, mitigates inter-symbol interference (ISI), and provides richer temporal context.

[0069] The data flow of the Bi-LSTM layer is as follows Figure 2 As shown in Figure 1, the Bi-LSTM network includes a forward long short-term memory network unit, a reverse long short-term memory network unit, and a splicing operation layer;

[0070] The forward LSTM network unit and the reverse LSTM network unit process the feature sequence data in a forward and reverse order respectively to obtain a forward feature sequence h = {h0, h1, ..., h n} and reverse feature sequence h'={h'0,h'1,...,h' n}, the splicing operation layer performs a splicing operation on the forward feature sequence and the reverse feature sequence to obtain the compensated feature sequence data.

[0071] In addition to performance, it is also crucial to fully evaluate the complexity of the proposed method. According to the characteristics of the TCN architecture, the complexity of the structure is calculated as follows:

[0072]

[0073] Among them, n i Represents the characteristics of the original data, n f Indicates the number of filters, n k Indicates the kernel size. For TCN, the size of each filter is (n s -n k +1). The complexity of the convolutional network can be expressed in terms of the number of features, the number of filters, the kernel size, and the output length L after the convolutional layer. out In the above formula, the number of features in the first convolutional layer corresponds to the number of features in the original data. In subsequent layers, the number of features is updated according to the number of filters set in the previous convolutional layer. Each residual block contains two convolutional structures, each with a different number of filters and kernel size. To distinguish them, let n k and n f Represents the parameters of the first layer, n'k and n' f represents the parameters of the second layer. N represents the number of residual blocks. Therefore, the complexity of the proposed TFM neural network equalizer can be written as:

[0074]

[0075] Generally speaking, floating-point operations per second (FLOPS) is a measure of hardware complexity. Based on the parameterization outlined in this letter, the complexity of the proposed algorithm is calculated to be 4571 FLOPS.

[0076] In view of the limitation of hardware resources, the proposed signal nonlinear compensation model (i.e., TFM neural network model) is experimentally verified in a 1024 / 4096-QAM signal transmission system at a 4GBd rate. It is important to clarify that, Figure 4 The experimental setup and the designed TFM neural network model shown are independent of the signal transmission rate and exhibit wide applicability.

[0077] At the transmitter, a probabilistic shaping encoder uses a pseudo-random binary sequence as input and generates a corresponding amplitude sequence through a sophisticated bit-to-symbol mapping strategy. Pilot symbols are then added in the digital signal processing (DSP) module to enhance signal quality. The data is then resampled and pulse-shaped using a root-root cosine (RRC) finite impulse response (FIR) filter with a roll-off factor of 0.05 to ensure that the signal's spectral characteristics and time-domain waveform meet transmission requirements.

[0078] The processed digital signal is fed into an arbitrary waveform generator (AWG) with an 8-GSa / s sampling rate and 14-bit resolution, where it is converted into an analog signal through a digital-to-analog conversion (DAC) process. This analog signal drives four channels of electrical signals that are input to a Mach-Zehnder modulator (MZM), generating a high-quality 1024-QAM optical signal. After power amplification by an erbium-doped fiber amplifier (EDFA), the signal strength is significantly boosted. At this point, the input power into the 80 km standard single-mode fiber (SSMF) can be precisely controlled using a variable optical attenuator (VOA). The 4096-QAM signal exhibits optimal performance at an input power of approximately -4 dBm.

[0079] A second VOA is also included in the signal transmission path to adjust the optical signal-to-noise ratio (OSNR) to meet the needs of bit error rate measurement; a third VOA ensures stable optical power received by the coherent receiver. At the receiving end, an optical bandpass filter (OBPF) effectively suppresses the amplified spontaneous emission (ASE) noise introduced by the EDFA. The signal is then transmitted along with the lightwave from the local oscillator (LO) to the coherent receiver. The receiver is equipped with a narrow linewidth (less than 1kHz) laser set to a wavelength of 1550.112nm, two 90-degree optical hybrids, and four balanced photodetectors for efficient optical-to-electrical conversion.

[0080] The analog-to-digital conversion (ADC) process is performed by a 50-GSa / s real-time oscilloscope. The acquired data then undergoes a series of preprocessing steps in an offline DSP module, including traditional algorithms such as dispersion compensation, clock recovery, polarization demultiplexing, and carrier recovery to optimize signal quality. Finally, after fine-tuning and decision-making by the TFM NN equalizer, the original signal is successfully restored, verifying the effectiveness and reliability of the signal nonlinearity compensation model described in this paper.

[0081] In this paper, the performance of PS 1024 / 4096-QAM signals using ideal rate-adaptive forward error correction (FEC) coding is evaluated. For probabilistically shaped signals with varying entropy values, the normalized generalized mutual information (NGMI) can be used as a useful channel measurement metric. Regardless of the modulation format, a specific soft-decision forward error correction (SD-FEC) scheme can be evaluated. This means that for both 1024-QAM and 4096-QAM, and even other higher-order modulation schemes that may be adopted in the future, the NGMI can be calculated to predict and optimize their error correction capabilities and system throughput under a specific SD-FEC scheme.

[0082] Among them, the definition of NGMI is as follows:

[0083] NGMI=1-(H-GMI) / (4m)

[0084] For 1024-QAM, m is 5, and for 4096-QAM, m is 6. Using 20% ​​and 25% LDPC overhead, the corresponding NGMI thresholds are 0.881 and 0.778, respectively. Generally speaking, as long as the NGMI exceeds these thresholds, QAM signals can be transmitted error-free.

[0085] like Figure 5As shown in (a) and (c) of Figure 1, the NGMI (normalized generalized mutual information) performance shows a significant downward trend as the signal entropy value increases. However, thanks to the excellent performance of the proposed TFM NN equalizer, the NGMI performance is significantly improved and successfully reaches the required NGMI threshold. In order to intuitively demonstrate this improvement, Figure 5 Figures (a) and (c) show the constellation diagrams of 1024-QAM and 4096-QAM signals with and without the TFM NN equalizer, respectively. It can be clearly seen that with the help of the TFM NN equalizer, the originally blurred and scattered constellation points become converged and clear, which directly reflects the significant improvement in signal quality.

[0086] Likewise, observe Figure 4 We can also find similar performance improvement trends in (b) and (d) of Figure 1. These results show that the TFM NN equalizer has a significant effect on improving the transmission performance of high-order QAM signals.

[0087] Ultimately, for 1024-QAM signals, while maintaining the LDPC (Low-Density Parity-Check) encoded NGMI above the 20% threshold, we achieved a raw spectral efficiency (SE) of 16.190 bits / s / Hz using an entropy of 17.0 bits / symbol. For 4096-QAM signals, when the LDPC-encoded NGMI reached the 25% threshold, we achieved a raw SE of 21.188 bits / s / Hz using an entropy of 21.4 bits / symbol. These results not only validate the effectiveness of the TFM NN equalizer in improving the transmission performance of high-order QAM signals, but also provide strong support for the design of future high-speed, high-density fiber-optic communication systems.

[0088] In summary, this paper innovatively proposes a temporal feature memory (TFM) neural network for PS-PDM (Probability Shaping-Polarization Multiplexing) ultra-high-order QAM (Quadrature Amplitude Modulation) coherent optical transmission systems. Experimental verification using PS-PDM 1024 / 4096-QAM signals transmitted over 80 km of standard single-mode fiber (SSMF) fully demonstrates the advanced nature and practicality of this technology.

[0089] The TFM neural network equalizer designed in this paper cleverly integrates a temporal convolutional network feature extraction layer with a bidirectional long short-term memory network, significantly optimizing the latter's performance. This innovative design enables efficient transmission of PS-PDM 1024 / 4096-QAM signals over an 80 km SSMF, achieving raw spectral efficiencies of 16.190 bits / s / Hz and 21.188 bits / s / Hz, respectively.

[0090] Compared to traditional neural network-based approaches with a linear equalizer, the TFM neural network equalizer of this invention not only has a simpler structure but also demonstrates superior performance in compensating for the nonlinear distortion of ultra-high-order QAM signals. It also exhibits significant performance advantages over the classic Volterra equalizer.

[0091] Therefore, the proposed method provides an economical and efficient solution for the generation and transmission of ultra-high-order signals in future commercial, high-capacity coherent optical transmission systems. This innovative achievement not only promotes the development of optical communication technology but also lays a solid foundation for building faster and more reliable fiber-optic communication networks.

[0092] Example 2

[0093] Based on the same inventive concept as that of the first embodiment, the present invention also provides a nonlinear compensation system for ultra-high-order QAM signals based on a time-feature memory neural network. Figure 6 As shown, the system includes the following modules:

[0094] The data acquisition and preprocessing module 100 is used to acquire multiple groups of quadrature amplitude modulation signals under different optical signal-to-noise ratios, perform zero-mean normalization processing on the quadrature amplitude modulation signals, and obtain preprocessed quadrature amplitude modulation signals;

[0095] A model building and training module 200 is used to divide the pre-processed quadrature amplitude modulation signal into a training set and a test set, build a signal nonlinear compensation model, and train the signal nonlinear compensation model using the training set to obtain a trained signal nonlinear compensation model;

[0096] The signal nonlinear compensation module 300 is configured to perform nonlinear compensation on the quadrature amplitude modulation signal to be compensated based on the trained signal nonlinear compensation model.

[0097] The present embodiment proposes a nonlinear compensation system for ultra-high-order QAM signals based on a time-feature memory neural network, which is used to implement the aforementioned nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network. Therefore, the specific implementation method of the nonlinear compensation system for ultra-high-order QAM signals based on a time-feature memory neural network can be found in the embodiment section of the aforementioned nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network. For example, the data acquisition and preprocessing module 100, the model construction and training module 200, and the signal nonlinear compensation module 300 are respectively used to implement steps S1, S2, and S3 of the nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network in the first embodiment. Therefore, its specific implementation method can refer to the description of the corresponding embodiments of each part. In order to avoid redundancy, it will not be repeated here.

[0098] Example 3

[0099] The present invention also provides a computer storage medium, which stores a computer software product. The computer software product includes several instructions for enabling a computer device to execute the instructions of the ultra-high-order QAM signal nonlinear compensation method based on time feature memory neural network described in Example 1.

[0100] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0101] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0102] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0104] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network, characterized in that: The following steps are involved: S1: Under different optical signal-to-noise ratios, obtaining multiple groups of quadrature amplitude modulation signals, performing zero-mean normalization processing on the quadrature amplitude modulation signals, and obtaining pre-processed quadrature amplitude modulation signals; S2: Dividing the preprocessed quadrature amplitude modulation signal into a training set and a test set, constructing a signal nonlinear compensation model, and training the signal nonlinear compensation model using the training set to obtain a trained signal nonlinear compensation model; S3: performing nonlinear compensation on the quadrature amplitude modulation signal to be compensated based on the trained signal nonlinear compensation model; The signal nonlinear compensation model includes an input layer, a temporal convolutional network layer, a bidirectional long short-term memory neural network layer and an output layer; The input layer processes the training set to obtain a modulated signal in the form of an embedded vector; The temporal convolutional network layer extracts features from the modulation signal in the embedded vector form to obtain feature sequence data; The bidirectional long short-term memory neural network layer processes the feature sequence data in a forward and reverse order respectively, captures the bidirectional time dependency therein, and obtains compensated feature sequence data; The output layer outputs the compensated feature sequence data; The temporal convolutional network layer is configured to include a plurality of residual units connected in sequence, wherein the residual unit includes a first branch and a second branch; the first branch includes an expanded convolutional layer, and the expression of its data processing process is as follows: Where s represents the element of the input sequence x, d is the dilation factor, k is the size of the filter kernel, sd·i represents the past direction, and f is the filter set; The second branch is a residual skip branch including a 1×1 point-by-point convolutional layer.

2. The nonlinear compensation method for ultra-high-order QAM signals based on a time-characteristic memory neural network according to claim 1, characterized in that: The first branch and the second branch respectively extract features of the modulated signal in the embedded vector form, and perform a splicing operation on the output features of the two branches to obtain the feature sequence data.

3. The nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network according to claim 1, characterized in that: The first branch also includes a batch normalization layer, an activation layer, a Dropout layer, a dilated convolutional layer, a batch normalization layer, an activation layer, and a Dropout layer connected in sequence.

4. The nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network according to claim 3, characterized in that: The Tanh function used by the activation layer.

5. The nonlinear compensation method for ultra-high-order QAM signals based on a time-feature memory neural network according to claim 1, characterized in that: The bidirectional long short-term memory neural network layer includes a forward long short-term memory network unit, a reverse long short-term memory network unit and a splicing operation layer; Among them, the forward long short-term memory network unit and the reverse long short-term memory network unit process the feature sequence data in forward and reverse order respectively to obtain a forward feature sequence and a reverse feature sequence, and the splicing operation layer performs a splicing operation on the forward feature sequence and the reverse feature sequence to obtain the compensated feature sequence data.

6. A nonlinear compensation system for ultra-high-order QAM signals based on a time-feature memory neural network, characterized in that: For implementing the method according to claim 1, the system comprises the following modules: A data acquisition and preprocessing module is used to acquire multiple groups of orthogonal amplitude modulation signals under different optical signal-to-noise ratios, perform zero-mean normalization processing on the orthogonal amplitude modulation signals, and obtain preprocessed orthogonal amplitude modulation signals; A model construction and training module is used to divide the preprocessed orthogonal amplitude modulation signal into a training set and a test set, construct a signal nonlinear compensation model, and train the signal nonlinear compensation model using the training set to obtain a trained signal nonlinear compensation model; The signal nonlinear compensation module is used to perform nonlinear compensation on the quadrature amplitude modulation signal to be compensated based on the trained signal nonlinear compensation model.

7. A computer storage medium, characterized in that The computer storage medium stores a computer software product, which includes several instructions for enabling a computer device to execute the ultra-high-order QAM signal nonlinear compensation method based on time characteristic memory neural network as described in any one of claims 1 to 5.

Images